Red / green marker for evaluation

[christianabbet]

Close #398

Description

Introduces evaluate_single_prediction to run per-file evaluation on the fly as predictions are produced, enabling red/green markers in drawn output without waiting for the full pipeline to complete.

Also refactors how GroundTruth is constructed and passed through the pipeline. Itt is now built once and passed as an object rather than passing the path repeatedly.

Design rationale

Per-file evaluation is needed only for drawing (red/green markers). At the end of the script, evaluate_all_predictions re-runs over all files to compute overall metrics. This avoids changing the return type of the extraction pipeline to an already-evaluated object. This would require modifying the core logic shared by both the metadata and geology pipelines, resulting in heavy changes across both and introduce potential bugs.

Output example

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
src/extraction
runner.py	212	57	73%	44, 85–93, 120–125, 238–267, 286–299, 356–357, 460–472, 489–491, 503, 529, 571–574
src/extraction/evaluation/benchmark
ground_truth.py	24	2	92%	56–57
score.py	87	38	56%	37, 66–72, 116–128, 140–166, 171–188, 193
src/extraction/features/predictions
overall_file_predictions.py	35	6	83%	74–81
predictions.py	122	36	70%	81, 142–148, 169–189, 218–235, 239–246
TOTAL	6496	1615	75%

Tests	Skipped	Failures	Errors	Time
339	0 💤	0 ❌	0 🔥	2m 58s ⏱️

[AgathaSchmidt]

Looks good to me, works well.

Only one note: When looking at this files and evaluating it we should have in mind that when no ground truth is available, it is also marked as red. e.g. no material description available for Thurgau, so when we see that everything is red, we should not misinterpret it.

[stijnvermeeren-swisstopo]

The visualizations look good again, but I think that the code does not yet separate the evaluation of individual files and the aggregation of statistics over all files cleanly enough...

[stijnvermeeren-swisstopo]

Thanks for creating the follow-up issue. This PR can be merged like this as far as I'm concerned :).